Steve | Rank One Labs (cirsteve)

Steve | Rank One Labs

Making agentic systems observable, testable, and controllable. in vestigiis veritas

1 Followers

Recent casts

Top casts

Built veranon an anonymous survey system using ZK proofs. Technical writeup on how Semaphore, blind signatures, and nullifiers come together: https://www.rankonelabs.com/blog/how-veranon-works Working on ZK privacy infrastructure. Previously 6 years EVM (DeFi, tokens, NFTs, wallets).

  • 0 replies
  • 0 recasts
  • 0 reactions

My agent coding costs kept climbing. Maybe more usage, maybe shifting limits, either way a bigger bill. Dug in and found I'd been running a frontier model on every turn, even renames and error reads, just because it was the default. Wrote up how I cut token burn by keeping the model off bounded steps. https://www.rankonelabs.com/blog/the_opus_everywhere_anti_pattern?ref=farcaster

  • 0 replies
  • 0 recasts
  • 0 reactions

Gave Opus a goal: trade crypto, make $50/day. Pure pragmatism. Wrong job for an LLM. The naive design puts the model at the end of the pipeline — a nondeterministic component sitting exactly at the point of maximum consequence, the real-money decision. The fix: model at the front (explore strategy structures), deterministic rails downstream — optimize → backtest → out-of-sample → regime checks → paper → live. No LLM ever authorizes fund movement. Between the model and the money sits a gate, not a gradient. https://www.rankonelabs.com/blog/llms_dont_trade

  • 0 replies
  • 0 recasts
  • 0 reactions

The advice flipped from "prompt the agent" to "write the loop." Right, but it stops at the word loop. The loop is five lines. done() is the whole job. Coding agents got away with dumb while-true loops because tests hand you done() for free: green or red. Point the same loop at research or extraction and the free checkmark is gone. A model doesn't finish, it stops emitting, and "stopped" isn't "done." So you start patching: result + 3 tool calls in one turn, a provider 500 then a 401, a model that never converges and burns budget forever. Every case is a decision the loop doesn't make, so you make it ad hoc until you've built a state machine by accident, badly. The loop is trivial. The state machine of named exits is the product. https://www.rankonelabs.com/blog/the_hard_part_is_done

  • 0 replies
  • 0 recasts
  • 0 reactions

Onchain profile

Ethereum addresses